Teng Xinzhi, Chen Yingxuan, Zhang Yawei, Ren Lei
Duke Kunshan University, Kunshan, China.
Department of Radiation Oncology, Duke University Medical Center, Durham, NC, USA.
Quant Imaging Med Surg. 2021 Feb;11(2):737-748. doi: 10.21037/qims-19-1058.
To investigate the feasibility of using a supervised convolutional neural network (CNN) to register phase-to-phase deformable vector field of lung 4D-CT/4D-cone beam CT for 4D dose accumulation, contour propagation, motion modeling, or target verification.
We built a CNN-based deep learning method to register the deformation field directly between phases of patients' 4D-CT or 4D-cone beam CT. The input consists of patch pairs of two phases, while the output is the corresponding deformation field that registers the patch pairs. The centers of the patch pairs were uniformly sampled across the lung, and the size of the patches was chosen to cover the range of the respiratory motion. The network was trained to generate deformation field that matches with the reference deformation field generated by VelocityAI (Varian). The network is structured with four convolutional layers, two average pooling layers, and two fully connected layers. Half mean squared error is applied to guide the study as loss function. Nine patients with eleven sets of 4D-CT/cone beam CT image volumes were used for training and testing. The performance of the network was validated with intra-patient and inter-patient setups.
Registered images were generated with Velocity deformation field and the CNN deformation field, respectively. Main anatomic features such as the main vessels and the diaphragm matched well between two deformed images. In the diaphragm region, the coefficients of cross-correlation, root mean squared error, and structural similarity index measure (SSIM) between deformed images registered by CNN and VelocityAI was calculated. The cross-correlation was above 0.9 for all the intra-patient cases.
Patch-based deep learning methods achieved comparable deformable registration accuracy as VelocityAI. Compared to VelocityAI, the deep learning method is fully automatic and faster without user dependency, which makes it more preferable in clinical applications.
研究使用监督式卷积神经网络（CNN）对肺部4D-CT/4D锥形束CT的逐相可变形矢量场进行配准，以用于4D剂量累积、轮廓传播、运动建模或靶区验证的可行性。
我们构建了一种基于CNN的深度学习方法，用于直接在患者的4D-CT或4D锥形束CT的各相之间配准变形场。输入由两个相的图像块对组成，而输出是配准图像块对的相应变形场。图像块对的中心在肺部均匀采样，图像块的大小选择为覆盖呼吸运动范围。该网络经过训练，以生成与VelocityAI（瓦里安）生成的参考变形场相匹配的变形场。该网络由四个卷积层、两个平均池化层和两个全连接层构成。应用半均方误差作为损失函数来指导研究。使用九名患者的十一组4D-CT/锥形束CT图像体积进行训练和测试。通过患者内和患者间设置验证了该网络的性能。
分别使用Velocity变形场和CNN变形场生成了配准图像。两个变形图像之间的主要解剖特征，如主要血管和膈肌，匹配良好。在膈肌区域，计算了由CNN和VelocityAI配准的变形图像之间的互相关系数、均方根误差和结构相似性指数测量（SSIM）。所有患者内病例的互相关均高于0.9。
基于图像块的深度学习方法实现了与VelocityAI相当的可变形配准精度。与VelocityAI相比，深度学习方法是完全自动的，速度更快，且不依赖用户，这使其在临床应用中更具优势。
